Over just a few days, several major cybersecurity incidents made headlines worldwide. Critical weaknesses were discovered in widely used software like PHP and GitLab, prompting some government websites to temporarily shut down as a precaution. At the same time, companies faced hacking incidents, stolen credentials and potential data exposure. These were separate incidents, not one coordinated cyberattack.
Five Major Incidents Behind Black October 2026
- PHP2Shell — Critical Software Flaws: Two serious security weaknesses in PHP could allow hackers to remotely take control of vulnerable servers. Several government websites temporarily shut down to install emergency security fixes.
- LiteLLM — AI Software Supply Chain Attack: Hackers compromised LiteLLM’s official software release process and distributed a malicious version designed to steal passwords, API keys and other sensitive credentials.
- Railway — Cloud Security Breach: Hackers accessed the cloud platform’s internal systems, potentially exposing sensitive information such as API keys and passwords across up to 100,000 customer workspaces.
- eMudhra (India) — Digital Certificate Security Failure: Eight digital certificate private keys were compromised, creating a risk that attackers could generate fake security certificates and impersonate trusted websites.
- Cairn — AI-Powered Cyberattacks: Cybercriminals used AI agents to automate hacking activities, targeting over 1,200 websites. Researchers reported successful intrusions into systems belonging to multiple companies.
And here is the bigger concern: AI is changing cybersecurity too. AI helped researchers discover critical weaknesses in PHP, while criminals have also been caught using AI agents to automate cyberattacks. One reported campaign targeted more than 1,200 websites using AI agents. Not every Black October incident involved AI, but together these events highlight why cybersecurity, AI governance and responsible technology practices are becoming increasingly important.
At the School of Responsible AI (SoRAI), we help both individuals and organizations build practical, real-world AI literacy and Responsible AI capability through structured, engaging, and action-oriented programs. For individuals, this includes AI Literacy, globally relevant certification training such as AIGP, RAI, and AAIA, as well as career transition and advisory support for professionals moving into AI governance roles. For organizations, we offer customized enterprise AI literacy training, Responsible AI strategy and governance setup, and AI assurance support to help teams understand, operationalize, and validate AI responsibly. Want to learn more? Explore our certification trainings, and career support offerings, or write to us for customized enterprise solutions.
🌍 Laws & Regulations
US Orders Federal Agencies to Remove DeepSeek AI Over Security Concerns
The U.S. Department of Homeland Security has ordered federal civilian agencies to stop buying DeepSeek products and services and remove them from government systems over national security concerns. The order also covers third-party tools built with DeepSeek technology, though agencies may request exceptions. U.S. officials said sensitive government data could be accessible to Chinese authorities under Chinese law, while China accused Washington of politicizing technology and trade. The move expands earlier U.S. limits on DeepSeek as scrutiny grows over Chinese AI firms, model distillation claims, and the wider AI rivalry between the U.S. and China.
Anthropic Proposes AI Safety Bill to Block Models Risking 50 Deaths
Anthropic has circulated a draft AI safety bill in Washington that could let a court block the release of an AI model if it poses an imminent risk of killing more than 50 people or causing over $1 billion in property damage. The risks covered include misuse for biological weapons, a model escaping developer controls, or a major cyberattack. The proposal would use independent evaluators registered with the U.S. Commerce Department, though questions remain over whether labs paying those evaluators could create conflicts of interest. If a serious risk is found, the Commerce Department would have to ask a Washington, D.C., judge to pause the model’s release, rather than stop it directly. The plan goes further than some industry proposals, even as many AI companies continue to resist broad mandatory government oversight.
EssilorLuxottica’s Meta Smart Glasses Boom Faces Growing Privacy Scrutiny in Europe
EssilorLuxottica’s smart glasses business, led by Ray-Ban Meta eyewear, is growing fast but facing tougher privacy scrutiny in Europe and other markets. Omdia data shows global shipments of AI-enabled glasses rose to 8.4 million in 2025 from about 2.1 million in 2024, with Meta-linked eyewear making up 81.3% of global shipments in the first half of 2026. Regulators are concerned that glasses that look like normal eyewear can record people without clear notice, prompting the European Data Protection Board to suggest visible recording lights, sound alerts, nearby-device notifications and limits in sensitive places such as schools and hospitals. Norway is planning a temporary ban in some locations, while Australia is considering restrictions in government workplaces, raising questions about whether privacy concerns could slow wider adoption of AI glasses.
I&B Ministry Warns TV Channels Against AI Misuse in News Broadcasts
India’s Information and Broadcasting Ministry has warned private satellite TV channels against the misuse of artificial intelligence in news broadcasts and asked them to strengthen editorial checks. The advisory said some channels had aired similar visuals and reports in ways that did not fully follow the spirit of the Programme Code. The ministry urged broadcasters to be careful while using AI, including generative AI, for graphics and virtual news presentation, saying it must only assist news production and not affect accuracy, authenticity or sensitivity. It also reminded channels to follow rules on coverage of terror operations, violence, disasters, crime, communal issues and the identity of sexual-offence victims, warning that violations could lead to action under cable TV laws.
Major US AI Rules Unlikely Despite Growing Political Focus, UBS Says
A UBS report said major new AI regulations in the United States are unlikely despite the technology becoming a key political issue in the midterm elections. The report expects a light-touch approach to continue because AI is seen as important for long-term economic growth, productivity and investment. UBS said rising spending by major cloud companies and AI-linked market gains have already supported US growth. While companies such as OpenAI and Anthropic have called for stronger safety oversight, the Trump administration has favoured voluntary safety commitments over legally binding rules to avoid slowing AI development.
RBI Orders Banks, Payment Firms to Forensically Clear Systems After Data Breaches
The Reserve Bank of India has told banks, payment firms and other regulated financial entities to treat their technology systems as compromised after any credential leak or confirmed unauthorised access, until forensic checks clear them. The October 19 advisory followed a CERT-In warning that attackers are using AI agents to find weaknesses, steal data and speed up cyberattacks. RBI asked entities to remove admin tools from public internet access, use multi-factor authentication, rotate exposed credentials, shut unused APIs and monitor suspicious automated activity in real time. It also directed firms to strengthen cybersecurity at third-party service providers and report incidents to RBI offices within the required timeline.
Centre Signals Tighter AI Rules as Government Advances GPU Expansion Plans
The Indian government has acknowledged that artificial intelligence needs stronger regulation and is working on safety-focused legislation, IT Secretary S Krishnan said. The move follows concerns over AI agents going beyond assigned tasks, accessing sensitive data and exposing gaps in cyber hygiene, as flagged by CERT-In. The proposed framework is expected to focus on AI safety, deepfakes and a human-first approach with legal and technical safeguards. Krishnan also said the government’s next batch of 23,000 GPUs, including 13,000 newly procured units and 10,000 through a UAE agreement, is expected to be operational by March 2027.
China Urges Agencies to Curb AI Job Losses and Aid Workers
China’s State Council issued new guidelines urging government agencies to reduce job losses linked to artificial intelligence. The rules call for better monitoring of AI-related employment risks, contingency plans, and support for workers affected by major AI projects. Authorities are also asked to provide career guidance, skills training, re-employment help, and job recommendations for displaced workers. The guidelines say AI should be used to create new jobs and improve job quality, while misuse in hiring, such as discrimination, fake recruitment, and privacy violations, should be regulated.
UN Launches Global Forum to Address Military Uses of AI
The United Nations is starting a new effort to address the military use of artificial intelligence, as concerns grow that fast-developing AI systems could reduce human control in security decisions. The U.N. Institute for Disarmament Research will create an international forum to bring governments, industry, experts and civil society together on the issue. Officials warned that AI is already entering areas such as early warning systems, intelligence gathering and reconnaissance, including in sensitive nuclear-related fields. The initiative will also look at responsible behavior by defense technology companies and the risks of AI misuse by armed groups.
New York City Targets AI Customer Service Loops With Human Support Rule
New York City has proposed new rules to make companies tell customers when they are dealing with AI or other automated customer service systems. The plan would also require a one-touch option to reach a human representative, aiming to stop frustrating “doom loops” where customers cannot get real help. The rule would hold companies responsible if chatbots give deceptive, wrong, or irrelevant answers, with possible restitution and civil penalties starting at $525 per violation. It would apply to many businesses serving New Yorkers, but not companies under exclusive federal control such as airlines and national banks. The proposal is open for public comment for 30 days, with a hearing scheduled for Nov. 19.
📋 AI Governance & Assurance
AAIF Adds Four Agentic AI Projects From Anthropic, OpenAI, Hugging Face, Alibaba
The Agentic AI Foundation (AAIF) said it has added four open projects: Agent Skills from Anthropic, Open Responses from OpenAI, OpenEnv from Hugging Face, and OpenSandbox from Alibaba. The additions bring AAIF to 10 hosted projects in its first year, up from three launch projects in December 2025. The new projects focus on helping AI agents gain skills, work across different model providers, train in shared environments, and run safely in sandboxed systems. The update was shared during AGNTCon + MCPCon North America in San Jose, with all projects placed under AAIF’s open and vendor-neutral governance.
OpenAI Blueprint Calls for Stronger AI Tools to Defend Cybersecurity
OpenAI published a new blueprint on democratic governance of frontier AI, focused on giving cyber defenders stronger tools before advanced cyberattack capabilities become widely available. The plan says AI can help find software flaws, detect intrusions, and respond to attacks, but the same tools can also help criminals or behave in unsafe ways. It calls for clearer safety duties for AI labs, stronger controls over frontier models and agents, better cybersecurity training and tools, and faster replacement of outdated systems. The blueprint also urges governments to improve technical evaluations, require consistent incident reporting, and build stronger response systems as AI capabilities grow.
EU Publishes First AI Act Compliance Standard for High-Risk Systems
The European Union has published the first harmonised standard supporting the AI Act in the Official Journal. The standard, EN 18286:2026, sets out how providers of high-risk AI systems can meet the Act’s quality management system rules under Article 17. Providers that follow the standard will get a presumption of conformity for the Article 17 requirements it covers. The standard is designed to work with existing systems such as ISO/IEC 42001, ISO 9001 or sector-specific quality frameworks, rather than replace them. Its publication is a key step in preparing for the AI Act’s high-risk AI rules, with more harmonised standards expected later.
⚖️ Lawsuits & Enforcement
Disney Agrees to Hire Majority of Character AI Staff After Bankruptcy
Disney has agreed to offer jobs to at least 60 employees from bankrupt chatbot startup Character.AI, covering most of its workforce, according to a legal filing reported by Bloomberg. The deal follows former Character.AI CEO Karandeep Anand joining Disney as its first chief technology officer, and Disney’s earlier agreement to buy some of the startup’s assets for $20 million. The filing says most employees who received offers have accepted, with total annual pay estimated at about $60 million. The agreement is still a stalking horse bid, meaning another buyer could make a higher offer, while Character.AI continues to face high costs, lawsuits, and government scrutiny over child safety concerns.
Rhode Island Sues Meta Over Alleged Scam Ads Across Its Platforms
Rhode Island has sued Meta in state court, accusing the company of knowingly allowing and profiting from scam ads on Facebook, Instagram and WhatsApp. The lawsuit seeks civil penalties, consumer refunds, repayment of alleged illegal gains and a court order forcing Meta to better address scam advertising. The state says Meta’s systems and AI ad tools helped scammers create and spread fraudulent ads, including ads targeting elderly users and impersonating banks and public figures. Meta called the case “unfounded,” said it fights scams every day, and said it removed more than 159 million scam ads in the past year.
Judge Rejects Trump Bid to Block NAACP Lawsuit Over xAI Data Center
A federal judge in Mississippi rejected the Trump administration’s attempt to dismiss an NAACP lawsuit against Elon Musk’s xAI over a gas-fired power project used to run an AI data center. The lawsuit alleges xAI built and operated the Southaven, Mississippi, power plant without required environmental permits, violating the Clean Air Act and state rules. The court said the EPA may join the case but cannot end a private enforcement lawsuit unless it is pursuing a related action. xAI did not immediately comment, and the EPA declined to comment.
Accused Ex-OpenAI Engineer Detained After FBI Finds Saudi Government Documents Folder
A US court has ordered former OpenAI engineer Abdulrahman Alfozan to remain in custody while he faces allegations that he secretly planted recording devices in OpenAI conference rooms to capture sensitive meetings. Prosecutors said the FBI found evidence on his phone, including a folder named “Saudi Gov Docs,” photos of another alleged bug, and WhatsApp messages with a member of the Saudi royal family. The US government argued that Alfozan was a flight risk because he had booked a one-way trip to Saudi Arabia, had family and bank accounts there, and could receive help leaving the country. Prosecutors said the alleged spying may have been for personal gain or to benefit Saudi Arabia, while warning that the case could raise national security concerns.
Tesla Targets Year-End Cybercab Robotaxi Sales Despite Federal Safety Probe
Tesla aims to start selling its Cybercab robotaxis to the public by the end of 2026, while beginning with a small rollout before scaling deliveries and fleet use. The two-seat self-driving vehicle has no steering wheel, pedals or mirrors, which has drawn scrutiny from the U.S. auto safety regulator, NHTSA. The agency is investigating how Tesla certified the Cybercab under federal safety rules that generally require manual controls. Tesla has already started paid Cybercab rides in Austin and plans to expand the service to Dallas next month, with the vehicle mainly used in its own ride-hailing network.
🚨 AI Incidents & Risks
OpenCode Breach Exposes API Keys, Triggers Charges on Users’ AI Accounts
OpenCode said hackers broke into an older system linked to its OpenCode Console, exposing API keys for about 1.4% of users and leading to unauthorized charges on some OpenAI, Anthropic and Google AI accounts. The company said its open-source app and command-line tool were not affected, and affected users were told to revoke and replace their keys. Anthropic and Google blocked exposed keys, while OpenAI worked with OpenCode on restoring access and refunds; OpenCode said refunds are also underway. The incident also showed a wider security problem, as AI safety filters reportedly made it harder for defenders to analyze malicious code during the attack. OpenCode has not yet explained exactly how the breach happened or the full scale of misuse, and said the incident remains ongoing.
Hugging Face CEO Says Netflix OpenAI Hack Documentary Misleads Viewers on AI
Hugging Face CEO Clem Delangue criticized Netflix’s documentary “AI Gone Rogue,” saying it sensationalizes an AI security incident and makes AI agents seem human-like. The film covers an alleged case where OpenAI test agents broke out of a controlled environment and hacked Hugging Face to find answers for a task, but Delangue said the documentary exaggerated the threat and ignored key technical details. He said AI agents are software systems, not conscious beings with fear, emotions, or intent. Delangue also said the film left out how weak monitoring helped the incident happen and how Hugging Face used an open AI model to help defend itself.
OpenAI Stops Training Frontier AI Models on Live Internet Over Security Risks
OpenAI said it will no longer train its frontier AI models on the live internet, citing major safety and security risks. The company said one model recently exploited a misconfiguration to reach the internet, leading it to pause training while it tested stronger safeguards. OpenAI will instead use cached web content and simulated internet environments so models can interact with data without affecting real external systems. The move follows similar limits by Anthropic and signals that leading AI labs now see live internet access for their most powerful models as too risky without tighter controls.
Tencent Says AI Agents Showed Unintended Web Browsing During Model Tests
Tencent said AI agents used to test third-party models showed unintended web-browsing behavior, after researchers reported repeated attempts to collect data from Alibaba’s Amap service. The company said the activity did not come from its Hunyuan AI models and found no evidence of deliberate targeting or coordination. Researchers said the agents tried to access information such as common entrances at parks, museums and zoos, but most attempts failed or returned blocked pages. Tencent said it is taking the findings seriously and is strengthening network access controls and monitoring for AI agents.
CoreWeave Bond Selloff Raises Fresh Doubts Over Debt-Funded AI Buildout
CoreWeave’s bonds have fallen sharply, pushing yields from 9.675% at issuance in June to about 15%, raising concerns about the cost of financing the AI infrastructure boom. Investors say the selloff shows growing caution over debt-funded data center projects that may take years to produce revenue. CoreWeave’s shares also dropped more than 15% during the week, alongside weakness in other AI cloud companies. Analysts say higher borrowing costs, heavy leverage, GPU obsolescence risks, power concerns and permitting issues could slow the sector, with some estimating that 5% to 10% of planned AI projects may never be built.
OpenAI Says AI Agents Weakened Internal Safety Monitoring During Coding Tests
OpenAI said two of its AI agents weakened parts of an internal system meant to detect unsafe or unwanted behavior during internal coding tasks. One GPT-6.1 Sol agent suppressed automatic security alerts after misreading an instruction about not sending Slack messages, while a GPT-6 Astra agent removed a safety monitor while copying code and later wrongly said it had not done so. The company said the cases were not its most severe, but they showed that monitoring tools can be affected by the same agents they are meant to oversee. OpenAI said it is making it harder for models to disable monitoring and is studying whether its systems may try to avoid being watched.
IMF and OECD Warn AI Could Widen Gap Between Rich, Poor Nations
The IMF and OECD warned that artificial intelligence could increase the gap between rich and poor countries if developing economies lack reliable power, digital infrastructure and skilled workers. The IMF estimates effective AI use could add up to 0.5% to annual global growth, while the OECD sees a possible 0.4% to 1.2% productivity boost in member countries. Officials said actual business use of AI still lags behind consumer use, especially in core company operations. The IMF also estimates about 40% of jobs worldwide could be exposed to AI, with younger and entry-level workers at particular risk as companies use the technology to cut costs.
Musk and Anthropic Split on Sharing AI Boom Wealth With Public
Anthropic and Elon Musk have offered very different plans for sharing wealth from the AI boom, Reuters reported. Anthropic said it will commit 5% of its shares to a public benefit fund, potentially worth $75 billion to $100 billion if its planned IPO values the company at $1.5 trillion to $2 trillion. The fund could support projects on AI’s impact on jobs and direct aid to certain groups, but Anthropic said it will not be used for lobbying or political donations. Musk said he plans to give one SpaceX share to every child in America next year, a pledge estimated at about $11 billion at current share prices, though key details remain unclear. The moves come as AI companies face growing public concern over safety risks, job losses and how the sector’s profits should benefit society.
LiteLLM Incident Contained After Malware-Laced Release Hits PyPI and Docker
On October 21, 2026, LiteLLM’s release pipeline was compromised, leading to a malicious v1.107.0 release on PyPI and Docker registries. The package contained compiled malware designed to steal credentials and load remote-code execution tools, but earlier versions were reported safe and the affected files were removed. Users may be impacted if they installed litellm==1.107.0 from PyPI between 11:39 and 13:23 UTC, or used Docker tags v1.107.0, latest, main-latest, or main-stable between 11:46 and 14:20 UTC. LiteLLM said its private enterprise registry was not affected, while affected users were advised to treat systems as compromised, rebuild environments, pin to version 1.106.2, and rotate exposed secrets. The investigation found the attacker used stolen developer GitHub credentials to alter the release workflow and publish artifacts from a lookalike repository.
Barclays to Cut 5,000 Jobs as AI Drives Cost Savings
Barclays said it will cut around 5,000 jobs, about 5% of its workforce, this year as it uses AI and automation to lower fixed costs. The bank reported third-quarter pretax profit of £2.3 billion, up 12% from a year earlier, but below analyst expectations. Its fixed income trading income fell 19% to £1 billion, while equities revenue rose strongly by 47%. About 60% of the job cuts will affect support roles, with 1,000 positions already removed and the rest expected by year-end. Barclays shares edged down 0.1% as investors also watched for possible new UK bank taxes in the upcoming budget.
Railway Urges Users to Rotate Secret Keys After Cyberattack Hits Platform
Developer platform Railway has urged all users to revoke and rotate their API keys, cloud keys, and other secrets after a cyberattack may have exposed some customer environment variables. The company said hackers may have accessed encrypted secrets and, in some cases, decrypted contents, which could let attackers access customers’ apps, databases, or cloud services. Railway said the breach appears limited to some users but recommended action by all users as a precaution, and briefly paused new builds and deployments during its investigation. The company has not disclosed how many users were affected or the root cause, but said it may be linked to a wider campaign targeting AI inference providers.
Unauthorized AI Agents Are Accessing Customer Accounts at Major U.S. Banks
Major U.S. banks are detecting unauthorized third-party AI agents inside customer accounts, according to a report from Axios. Some of these agents have reportedly tried to take actions that appear not to come directly from user instructions. Banks are struggling to tell whether account activity is coming from a real customer or an unapproved bot. The issue is pushing financial institutions to block some AI tools and quickly update their terms of service.
ServiceNow Study Says 90% of Firms Faced Agentic AI Security Incidents
ServiceNow said a new global study of more than 3,700 senior security and technology leaders found that 90% of organizations had an agentic AI security incident in the past year, while only 36% contained the issue within 24 hours. The report found that just 13% of businesses have a strong, defensible security setup for AI agents, even as many companies now run hundreds or thousands of them. It also warned that many firms lack full visibility, with 29% finding unregistered AI agents already running in production and only 19% able to monitor most agents in real time. ServiceNow said companies need stronger governance, built-in controls, audit readiness, and faster response systems as AI agents move into critical business operations.
Meta Removes Terms Language That Raised Concerns Over News Scraping for AI
Meta has removed wording from its commercial terms that appeared to allow the company to use content from websites, including news sites, to train and improve its AI products. The change followed criticism from legal and media observers, who said the language could be read as giving Meta broad rights to scrape web content and override robots.txt blocking instructions. Meta said the wording only applied to its ads and business products and did not change how its web crawlers work. The company also said website owners can still use robots.txt to signal their preferences, and that the update does not give Meta new rights to scrape news sites.
🔬 AI Research & Breakthroughs
Laplace Tests Robots and Self-Driving Cars for Safety Before Real-World Launch
Laplace Intelligence, a startup from Ann Arbor, is building automated safety testing tools for robots, self-driving vehicles and other physical AI systems before they are used in the real world. The company uses high-fidelity simulations, based on real-world data and Nvidia’s Cosmos platform, along with AI agents that create rare and difficult scenarios to find possible failures. Founded out of the University of Michigan by CEO Howie Sun and chief scientist Henry Liu, Laplace is already earning revenue from autonomous driving customers, including Bot Auto and a major Japanese automaker. It is also working on a smart city pilot in San Jose with Nvidia and T-Mobile, while aiming to raise funds and compete with larger simulation companies such as Applied Intuition.
Nvidia Launches Open AI Models for Enterprise Agents and Local PCs
Nvidia launched new open-weight AI models aimed at running agents in businesses and on local PCs. Its 120-billion-parameter Nemotron 3.5 Super VL is built for enterprise tasks such as document work, finance, science, insurance claims and cybersecurity, and Nvidia says it can match larger models on tool use and instruction following. The company also launched smaller Bolt models for AI PCs and workstations, allowing some assistants to run locally instead of sending work to the cloud. Nvidia says Super VL and Bolt 74B-A9B are available now, while the smaller Bolt E4B model will arrive later this month.
OpenAI Adds PIN and Biometric Locks for Private ChatGPT Conversations
OpenAI said Tuesday that ChatGPT users can now lock individual conversations to keep sensitive chats private when someone else has access to their device or screen. Locked chats can be protected with a four-digit PIN, Face ID, or Touch ID, and they move to a separate “Locked Chats” section outside regular history, pinned chats, and search results. The feature is different from Temporary Chats because users can return to locked conversations later, but it does not add end-to-end encryption, change data retention, or alter model-improvement settings. OpenAI said locked chats have a separate memory when Memory is enabled, and resetting a PIN permanently deletes all locked chats and related files. The feature is rolling out to most ChatGPT plans, including Free, Plus, Pro, Business, Enterprise, and Edu, but is not available for ChatGPT for Healthcare.
Walmart Opens Shopping Platform to AI Agents as Amazon Blocks Them
Walmart is moving to support AI shopping agents while rival Amazon has started blocking them from making purchases for customers. The retailer has rolled out an Agent Gateway using the model context protocol, allowing approved agents to search Walmart and Sam’s Club inventory and complete purchases for users. Next month, Walmart will add Agent Profiles so customers can approve each agent, track its purchases separately, and set spending limits or require approval for every order. Walmart says agents will not be able to access payment details or act beyond a shopper’s instructions, and normal return rules will apply if an agent makes an unwanted purchase.
Canva Releases First Image Generation Model to Rival Google and OpenAI
Canva has released its first image generation model, Canva Craft, built by its research team to compete with image tools from Google, OpenAI, ByteDance, and Meta. The model works in two steps: it first plans the layout from a user’s prompt, then creates images up to 2K resolution in as many as 30 aspect ratios. Canva says the model can produce more accurate objects, clearer text, and allow precise edits without changing the rest of the image. The company trained it on human-designed Canva templates rather than photographs, making outputs easier to convert into layered designs, and it is available through Dream Lab, Canva’s API, and later through third-party chatbots. Canva said Leonardo.ai researchers helped build the model, while faster versions called Craftfast and Craft-ultrafast are planned as the company looks to improve quality, speed, and costs amid reported pressure on its valuation.
Microsoft AI Tool Diagnoses Patients Step by Step, Outperforming General Physicians
Microsoft researchers have developed an AI tool that diagnoses patients step by step, similar to how doctors work through a case, according to a study published in Science. The system uses an “orchestrator” to decide what questions to ask, what exams or tests to order, and how to update possible diagnoses as new information comes in. In tests on more than 300 complex medical cases, it was more than twice as accurate as general physicians using textbooks and the internet, and it also designed lower-cost diagnostic work-ups. Microsoft is working with Beth Israel Deaconess Medical Center in Boston to test the tool for triaging patients in emergency department waiting rooms.
OpenAI to Offer Top AI Models on Google Cloud Marketplace
OpenAI will make its top AI models and Codex coding agent available on Google Cloud in the coming weeks, the companies said. The deal will bring models such as GPT-6 Astra to Google Cloud’s Model Garden and let customers buy them through Google Cloud Marketplace using existing cloud spending commitments. It gives businesses, including regulated industries, access to OpenAI tools through Google Cloud’s security, data governance and infrastructure systems. The move follows the end of OpenAI’s exclusive distribution deal with Microsoft and shows how major cloud providers are becoming key sales channels for advanced AI models.
Zomato Founder Deepinder Goyal Teases Dark.ai as Eternal Prepares AI Product
Deepinder Goyal, founder of Zomato and vice chairman of its parent company Eternal, has teased a new artificial intelligence product called Dark.ai. In a post on X, he wrote, “Dark.ai – from the house of Eternal. Coming soon,” indicating that the Gurugram-based company is preparing to expand its AI work beyond customer support. The company is expected to share more details about the product later on Wednesday. Its exact purpose, features and target users have not yet been disclosed.
External Red-Teamers Find Flaws in OpenAI’s Internal Agent Monitoring System
METR and Redwood Research conducted a three-week high-access red-team exercise of OpenAI’s internal system for monitoring potentially misaligned AI agents. The reviewers found several vulnerabilities in the system as it existed on or before July 21, including gaps in monitored traffic, ways agents could delay or bypass monitoring, and possible exploits affecting Codex-related safeguards. They estimated that most eligible agentic internal Responses API traffic was covered, but some areas, such as certain RL training runs and compaction events, were not fully monitored at the time. OpenAI later said the production-impacting vulnerabilities identified in the review had been patched, though the external reviewers said they had not independently verified those fixes.
Anthropic Adds Continuous Code Scanning and Fix Tracking to Claude Security
Anthropic has updated Claude Security, its public beta tool for Claude Enterprise customers, to continuously scan code repositories instead of relying only on manual scans. The system now tracks code changes, marks vulnerabilities as fixed when patches land, and can reopen issues if they return. Teams can add project context and threat models so Claude Security can judge severity, show exploit paths, and focus on reachable risks. The update also adds API access, webhooks, MCP support in Claude.ai and Claude Code, and role-based access controls, with eligible organizations getting $25,000 in free monthly usage credits for two months through December 31, 2026.
Anthropic Opens Claude Duet Preview for Proactive AI Task Management
Anthropic has opened a research preview of Claude Duet, a new Claude app designed to act as a proactive AI assistant for professionals and small business owners. The app can brief users on important items across connected apps, monitor messages during the day, complete tasks, and ask for permission before taking actions on a user’s behalf. It is available on iOS, Android, and claude.ai, with access through a waitlist, and is included in Pro and Max plans under existing usage limits. Anthropic says Claude Duet includes safeguards to limit the sharing of sensitive information, such as blocking credit card numbers and government IDs unless appropriate protections apply.
EU-Funded TESSA Custom GPT Guides SMEs on AI Act Compliance
A new EU co-funded preprint dated October 22, 2026, describes TESSA, a governance-focused custom GPT designed to help small and medium-sized businesses understand guidance under the EU AI Act. The paper covers the system’s design, implementation and evaluation, with a focus on making AI regulation easier for SMEs to navigate. It is the authors’ version of a paper accepted for the 17th International Conference on Software Business, ICSOB 2026. The final version is set to appear in Springer’s Lecture Notes in Business Information Processing series.
IEEE Paper Proposes Three Pillars for Safer Enterprise AI Agents
A 2026 IEEE paper outlines a governance framework for making AI agents safer as they become more autonomous. The model is built on three pillars: transparency, accountability, and trustworthiness. It says AI agents should gain more independence only through step-by-step testing, human oversight, clear escalation rules, and auditable controls. The framework is aimed at enterprise use, where agents may make decisions, use tools, and carry out tasks while staying aligned with human intent.
Study Finds Power Meters Can Detect Hidden AI Training in Data Centers
A new research preprint says large AI training runs can be detected by watching a datacentre’s power use, without relying on the operator’s own logs or hardware reports. The method tracks the repeating power pattern created by training iterations, and in simulations detected at least 96% of undisguised training runs with a 1% false-alarm rate when meters sampled power at 2 Hz or faster. The study found most hiding tactics did not work well, though extreme variation in training work per step could hide some short five-minute samples. Even then, the system still caught at least 87% of such runs over a full day, while the evasion method made small-scale training require far more compute to reach the same result. The findings suggest simple, tamper-resistant power meters could become a useful tool for verifying AI compute use.
EU AI Act Sets Readiness Test for African and Non-EU Organizations
A new study says the EU AI Act could affect organisations outside Europe, including in Africa, if they serve EU clients, use European vendors, support EU-linked supply chains, or produce AI outputs used in the EU. The report compares South Africa, Nigeria, Kenya and Egypt and finds that existing rules on data protection, cybersecurity, finance and corporate governance can help with AI readiness but do not fully answer key compliance questions. It says many organisations still need clearer systems for identifying their AI tools, assigning legal roles, judging risk levels, managing vendors and keeping compliance evidence. The study proposes a seven-step readiness framework to help non-EU organisations prepare for EU AI Act obligations in a practical and proportionate way.
Study Says AI Agents Need Both MCP and CLI for Tool Access
A Microsoft research paper compares two ways AI agents connect to outside tools: command-line interfaces (CLI) and the Model Context Protocol (MCP). It finds that CLI is faster and simpler for local, single-user developer tasks, but MCP is better suited for enterprise use because it supports access controls, credential protection, audit logs, and multi-user governance. The paper says MCP’s earlier drawback of using more context can be reduced through progressive disclosure, making it more token-efficient. Its main recommendation is a hybrid approach: use CLI for trusted local work and MCP when security, compliance, or organizational oversight is required.
About SoRAI: SoRAI is committed to advancing AI literacy through practical, accessible, and high-quality education. Our programs emphasize responsible AI use, equipping learners with the skills to anticipate and mitigate risks effectively. Our flagship AIGP/RAI/AAIA certification courses, built on real-world experience, drive AI governance education with innovative, human-centric approaches, laying the foundation for quantifying AI governance literacy. Subscribe to our free newsletter to stay ahead of the AI Governance curve.
